I can assure I work much faster. Maybe it is a bit different since I work in research and that’s indeed different from working on a large established codebase. Most of my projects are greenfield.
However, recently using Claude code I started many different projects I’d never have approached since I knew it would have taken me months to complete correctly. We are talking about porting file format readers and writers to new languages, then implementing novel algorithms to process such data and optimize it to work on different GPU architectures. Getting a working software takes about a week of work. A publishable cleaned up codebase 2/3 weeks. No way I could do this myself alone in such span of time.
On the other hand, I have students working with me, using AI is definitely not helping them to learn how to code and how to reason. In several occasions I had them showing me a novel huge equation which apparently worked, but then looking at it properly was just over fitting data and they had no way to explain why such an equation should be used.
For coding tasks I treat A.I. as a fresh intern that doesn’t really know what they’re doing but you can just ask them to do that horribly tedious task and just do a code review later. As you said, something like “rewrite this in Rust”, or “create unit tests for this function”. Then you fix all it’s mistakes. I’d never let it touch an actual algorithm or anything important though.
To be fair, it is fairly good at selecting and optimizing algorithms; as long as you set up a set of correctness tests and performance benchmarks it can use to evaluate if it’s doing well.
The other day in about half an hour I got a 26,000% improvement in performance on a specific algorithm which I had implemented as it was published in the original paper. Same results under all conditions, but it decided to rearrange the formula to use a few matrix tricks and optimized for SIMD instructions. That is the kind of things I’d never do myself: I don’t care if such calculation takes 3 weeks using 256 cores and 200 GB of memory, if I have to optimize for SIMD instructions that will likely take me one or two weeks and I don’t have time for that. I’d rather burn compute. If it takes me a couple hours I’m very happy to bring down a computation which would take weeks to an hour or so.
I can assure I work much faster. Maybe it is a bit different since I work in research and that’s indeed different from working on a large established codebase. Most of my projects are greenfield.
However, recently using Claude code I started many different projects I’d never have approached since I knew it would have taken me months to complete correctly. We are talking about porting file format readers and writers to new languages, then implementing novel algorithms to process such data and optimize it to work on different GPU architectures. Getting a working software takes about a week of work. A publishable cleaned up codebase 2/3 weeks. No way I could do this myself alone in such span of time.
On the other hand, I have students working with me, using AI is definitely not helping them to learn how to code and how to reason. In several occasions I had them showing me a novel huge equation which apparently worked, but then looking at it properly was just over fitting data and they had no way to explain why such an equation should be used.
For coding tasks I treat A.I. as a fresh intern that doesn’t really know what they’re doing but you can just ask them to do that horribly tedious task and just do a code review later. As you said, something like “rewrite this in Rust”, or “create unit tests for this function”. Then you fix all it’s mistakes. I’d never let it touch an actual algorithm or anything important though.
To be fair, it is fairly good at selecting and optimizing algorithms; as long as you set up a set of correctness tests and performance benchmarks it can use to evaluate if it’s doing well. The other day in about half an hour I got a 26,000% improvement in performance on a specific algorithm which I had implemented as it was published in the original paper. Same results under all conditions, but it decided to rearrange the formula to use a few matrix tricks and optimized for SIMD instructions. That is the kind of things I’d never do myself: I don’t care if such calculation takes 3 weeks using 256 cores and 200 GB of memory, if I have to optimize for SIMD instructions that will likely take me one or two weeks and I don’t have time for that. I’d rather burn compute. If it takes me a couple hours I’m very happy to bring down a computation which would take weeks to an hour or so.